GNGTS 2024 - Atti del 42° Convegno Nazionale

Session 3.3 GNGTS 2024 Ensemble-based Acoustc Full Waveform Inversion: A Synthetc Data Applicaton F. Macelloni 1 , M. H. Altaf 1 , M. Aleardi 1 , E.M. Stucchi 1 1 Department of Earth Sciences, University of Pisa, Pisa, Italy Introducton Full Waveform Inversion (FWI) is one of the most powerful techniques to estmate the distributon of seismic wave velocity in the subsurface. The determinaton of the velocites from the recorded seismograms represents an inverse problem and FWI aims to solve it by exploitng the full informaton content of the data. Despite the high resoluton results that FWI is able to provide, there are some drawbacks we have to deal with when using this kind of optmizaton. One of them is the risk of being trapped in local minima of the objectve functon, which expresses the distance between observed and estmated data. This problem is mainly due to the lack of low frequencies in the data (cycle skipping issue) and to a startng model lying too far from the global minimum of the error functon. To alleviate this problem, a global optmizaton approach could be adopted to replace the standard local, deterministc strategy, at the expense of a signifcant increase of the computatonal workload. Another limitaton of the deterministc inversion is also the impossibility to assess the uncertainty afectng the estmated subsurface velocity model. In this work we cast the FWI into a probabilistc framework. The aim of the work is twofold: making the FWI results less dependent from the startng model, while also estmatng the uncertainty on the inversion outcomes. Therefore, our aim is not to estmate a single, best-ftng soluton but providing as the fnal results the so called posterior probability density functon from which extract signifcant statstcal propertes concerning the estmated model (i.e., mean model and the associated standard deviaton). In partcular, we present an ensemble-based approach to FWI, using the Ensemble Smoother with Multple Data Assimilaton (ES-MDA) algorithm (Emerick et al., 2013). This method allows us to perform a Bayesian FWI by considering an ensemble of velocity models and iteratvely updatng each of these realizatons. The underlying assumpton is that data and model parameters follow a Gaussian distributon. MDA can be considered as an iteratve version of the standard ES and, instead of a single and large correcton, it performs multple smaller updates, achieving good data predictons in less iteratons. For additonal details, see Thurin et al. (2019) and Aleardi et al. (2021b).